xfs: remove wrapper for the fsync file operation
authorChristoph Hellwig <hch@infradead.org>
Mon, 15 Feb 2010 09:44:48 +0000 (09:44 +0000)
committerAlex Elder <aelder@sgi.com>
Mon, 1 Mar 2010 22:34:38 +0000 (16:34 -0600)
Currently the fsync file operation is divided into a low-level
routine doing all the work and one that implements the Linux file
operation and does minimal argument wrapping.  This is a leftover
from the days of the vnode operations layer and can be removed to
simplify the code a bit, as well as preparing for the implementation
of an optimized fdatasync which needs to look at the Linux inode
state.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
fs/xfs/linux-2.6/xfs_file.c
fs/xfs/xfs_vnodeops.c
fs/xfs/xfs_vnodeops.h

index 1eb561a..6c283b7 100644 (file)
@@ -35,6 +35,7 @@
 #include "xfs_dir2_sf.h"
 #include "xfs_dinode.h"
 #include "xfs_inode.h"
+#include "xfs_inode_item.h"
 #include "xfs_bmap.h"
 #include "xfs_error.h"
 #include "xfs_rw.h"
@@ -96,6 +97,120 @@ xfs_iozero(
        return (-status);
 }
 
+/*
+ * We ignore the datasync flag here because a datasync is effectively
+ * identical to an fsync. That is, datasync implies that we need to write
+ * only the metadata needed to be able to access the data that is written
+ * if we crash after the call completes. Hence if we are writing beyond
+ * EOF we have to log the inode size change as well, which makes it a
+ * full fsync. If we don't write beyond EOF, the inode core will be
+ * clean in memory and so we don't need to log the inode, just like
+ * fsync.
+ */
+STATIC int
+xfs_file_fsync(
+       struct file             *file,
+       struct dentry           *dentry,
+       int                     datasync)
+{
+       struct xfs_inode        *ip = XFS_I(dentry->d_inode);
+       struct xfs_trans        *tp;
+       int                     error = 0;
+       int                     log_flushed = 0;
+
+       xfs_itrace_entry(ip);
+
+       if (XFS_FORCED_SHUTDOWN(ip->i_mount))
+               return -XFS_ERROR(EIO);
+
+       xfs_iflags_clear(ip, XFS_ITRUNCATED);
+
+       /*
+        * We always need to make sure that the required inode state is safe on
+        * disk.  The inode might be clean but we still might need to force the
+        * log because of committed transactions that haven't hit the disk yet.
+        * Likewise, there could be unflushed non-transactional changes to the
+        * inode core that have to go to disk and this requires us to issue
+        * a synchronous transaction to capture these changes correctly.
+        *
+        * This code relies on the assumption that if the i_update_core field
+        * of the inode is clear and the inode is unpinned then it is clean
+        * and no action is required.
+        */
+       xfs_ilock(ip, XFS_ILOCK_SHARED);
+
+       if (ip->i_update_core) {
+               /*
+                * Kick off a transaction to log the inode core to get the
+                * updates.  The sync transaction will also force the log.
+                */
+               xfs_iunlock(ip, XFS_ILOCK_SHARED);
+               tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_FSYNC_TS);
+               error = xfs_trans_reserve(tp, 0,
+                               XFS_FSYNC_TS_LOG_RES(ip->i_mount), 0, 0, 0);
+               if (error) {
+                       xfs_trans_cancel(tp, 0);
+                       return -error;
+               }
+               xfs_ilock(ip, XFS_ILOCK_EXCL);
+
+               /*
+                * Note - it's possible that we might have pushed ourselves out
+                * of the way during trans_reserve which would flush the inode.
+                * But there's no guarantee that the inode buffer has actually
+                * gone out yet (it's delwri).  Plus the buffer could be pinned
+                * anyway if it's part of an inode in another recent
+                * transaction.  So we play it safe and fire off the
+                * transaction anyway.
+                */
+               xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
+               xfs_trans_ihold(tp, ip);
+               xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
+               xfs_trans_set_sync(tp);
+               error = _xfs_trans_commit(tp, 0, &log_flushed);
+
+               xfs_iunlock(ip, XFS_ILOCK_EXCL);
+       } else {
+               /*
+                * Timestamps/size haven't changed since last inode flush or
+                * inode transaction commit.  That means either nothing got
+                * written or a transaction committed which caught the updates.
+                * If the latter happened and the transaction hasn't hit the
+                * disk yet, the inode will be still be pinned.  If it is,
+                * force the log.
+                */
+               xfs_iunlock(ip, XFS_ILOCK_SHARED);
+               if (xfs_ipincount(ip)) {
+                       if (ip->i_itemp->ili_last_lsn) {
+                               error = _xfs_log_force_lsn(ip->i_mount,
+                                               ip->i_itemp->ili_last_lsn,
+                                               XFS_LOG_SYNC, &log_flushed);
+                       } else {
+                               error = _xfs_log_force(ip->i_mount,
+                                               XFS_LOG_SYNC, &log_flushed);
+                       }
+               }
+       }
+
+       if (ip->i_mount->m_flags & XFS_MOUNT_BARRIER) {
+               /*
+                * If the log write didn't issue an ordered tag we need
+                * to flush the disk cache for the data device now.
+                */
+               if (!log_flushed)
+                       xfs_blkdev_issue_flush(ip->i_mount->m_ddev_targp);
+
+               /*
+                * If this inode is on the RT dev we need to flush that
+                * cache as well.
+                */
+               if (XFS_IS_REALTIME_INODE(ip))
+                       xfs_blkdev_issue_flush(ip->i_mount->m_rtdev_targp);
+       }
+
+       return -error;
+}
+
 STATIC ssize_t
 xfs_file_aio_read(
        struct kiocb            *iocb,
@@ -755,7 +870,8 @@ write_retry:
                        mutex_lock(&inode->i_mutex);
                xfs_ilock(ip, iolock);
 
-               error2 = xfs_fsync(ip);
+               error2 = -xfs_file_fsync(file, file->f_path.dentry,
+                                        (file->f_flags & __O_SYNC) ? 0 : 1);
                if (!error)
                        error = error2;
        }
@@ -826,28 +942,6 @@ xfs_file_release(
        return -xfs_release(XFS_I(inode));
 }
 
-/*
- * We ignore the datasync flag here because a datasync is effectively
- * identical to an fsync. That is, datasync implies that we need to write
- * only the metadata needed to be able to access the data that is written
- * if we crash after the call completes. Hence if we are writing beyond
- * EOF we have to log the inode size change as well, which makes it a
- * full fsync. If we don't write beyond EOF, the inode core will be
- * clean in memory and so we don't need to log the inode, just like
- * fsync.
- */
-STATIC int
-xfs_file_fsync(
-       struct file             *file,
-       struct dentry           *dentry,
-       int                     datasync)
-{
-       struct xfs_inode        *ip = XFS_I(dentry->d_inode);
-
-       xfs_iflags_clear(ip, XFS_ITRUNCATED);
-       return -xfs_fsync(ip);
-}
-
 STATIC int
 xfs_file_readdir(
        struct file     *filp,
index ddd2c5d..9d376be 100644 (file)
@@ -583,113 +583,6 @@ xfs_readlink(
        return error;
 }
 
-/*
- * xfs_fsync
- *
- * This is called to sync the inode and its data out to disk.  We need to hold
- * the I/O lock while flushing the data, and the inode lock while flushing the
- * inode.  The inode lock CANNOT be held while flushing the data, so acquire
- * after we're done with that.
- */
-int
-xfs_fsync(
-       xfs_inode_t     *ip)
-{
-       xfs_trans_t     *tp;
-       int             error = 0;
-       int             log_flushed = 0;
-
-       xfs_itrace_entry(ip);
-
-       if (XFS_FORCED_SHUTDOWN(ip->i_mount))
-               return XFS_ERROR(EIO);
-
-       /*
-        * We always need to make sure that the required inode state is safe on
-        * disk.  The inode might be clean but we still might need to force the
-        * log because of committed transactions that haven't hit the disk yet.
-        * Likewise, there could be unflushed non-transactional changes to the
-        * inode core that have to go to disk and this requires us to issue
-        * a synchronous transaction to capture these changes correctly.
-        *
-        * This code relies on the assumption that if the update_* fields
-        * of the inode are clear and the inode is unpinned then it is clean
-        * and no action is required.
-        */
-       xfs_ilock(ip, XFS_ILOCK_SHARED);
-
-       if (!ip->i_update_core) {
-               /*
-                * Timestamps/size haven't changed since last inode flush or
-                * inode transaction commit.  That means either nothing got
-                * written or a transaction committed which caught the updates.
-                * If the latter happened and the transaction hasn't hit the
-                * disk yet, the inode will be still be pinned.  If it is,
-                * force the log.
-                */
-               xfs_iunlock(ip, XFS_ILOCK_SHARED);
-               if (xfs_ipincount(ip)) {
-                       if (ip->i_itemp->ili_last_lsn) {
-                               error = _xfs_log_force_lsn(ip->i_mount,
-                                               ip->i_itemp->ili_last_lsn,
-                                               XFS_LOG_SYNC, &log_flushed);
-                       } else {
-                               error = _xfs_log_force(ip->i_mount,
-                                               XFS_LOG_SYNC, &log_flushed);
-                       }
-               }
-       } else  {
-               /*
-                * Kick off a transaction to log the inode core to get the
-                * updates.  The sync transaction will also force the log.
-                */
-               xfs_iunlock(ip, XFS_ILOCK_SHARED);
-               tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_FSYNC_TS);
-               error = xfs_trans_reserve(tp, 0,
-                               XFS_FSYNC_TS_LOG_RES(ip->i_mount), 0, 0, 0);
-               if (error) {
-                       xfs_trans_cancel(tp, 0);
-                       return error;
-               }
-               xfs_ilock(ip, XFS_ILOCK_EXCL);
-
-               /*
-                * Note - it's possible that we might have pushed ourselves out
-                * of the way during trans_reserve which would flush the inode.
-                * But there's no guarantee that the inode buffer has actually
-                * gone out yet (it's delwri).  Plus the buffer could be pinned
-                * anyway if it's part of an inode in another recent
-                * transaction.  So we play it safe and fire off the
-                * transaction anyway.
-                */
-               xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
-               xfs_trans_ihold(tp, ip);
-               xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
-               xfs_trans_set_sync(tp);
-               error = _xfs_trans_commit(tp, 0, &log_flushed);
-
-               xfs_iunlock(ip, XFS_ILOCK_EXCL);
-       }
-
-       if (ip->i_mount->m_flags & XFS_MOUNT_BARRIER) {
-               /*
-                * If the log write didn't issue an ordered tag we need
-                * to flush the disk cache for the data device now.
-                */
-               if (!log_flushed)
-                       xfs_blkdev_issue_flush(ip->i_mount->m_ddev_targp);
-
-               /*
-                * If this inode is on the RT dev we need to flush that
-                * cache as well.
-                */
-               if (XFS_IS_REALTIME_INODE(ip))
-                       xfs_blkdev_issue_flush(ip->i_mount->m_rtdev_targp);
-       }
-
-       return error;
-}
-
 /*
  * Flags for xfs_free_eofblocks
  */
index ee33e11..36f3858 100644 (file)
@@ -21,7 +21,6 @@ int xfs_setattr(struct xfs_inode *ip, struct iattr *vap, int flags);
 #define XFS_ATTR_NOACL         0x08    /* Don't call xfs_acl_chmod */
 
 int xfs_readlink(struct xfs_inode *ip, char *link);
-int xfs_fsync(struct xfs_inode *ip);
 int xfs_release(struct xfs_inode *ip);
 int xfs_inactive(struct xfs_inode *ip);
 int xfs_lookup(struct xfs_inode *dp, struct xfs_name *name,